Physics-informed Machine Learning (PIML)-guided Contrails Formation Prediction

Contrail cirrus is estimated to be responsible for more than 50% of aviation induced climate forcing to date. However operational contrail avoidance is currently not possible due to the inability to predict exactly when and where a persistent contrail will form, and how that translates into radiative forcing. This research presents how to construct a high-accuracy physics-informed machine learning (PIML) models based on engine and weather variables to predict contrails formation parameters such as visibility, onset, and plumes’ optical depth as a function of distance from the aircraft. The approach is based on nonintrusive PIML model with low compute need at inference, making it ideal for onboard deployment. Based on a comprehensive sensitivity and feature importance study of the PIML model, this work demonstrates that spatial variation of plumes’ optical depth and as a result, the onset probability and location of contrails formation are only sensitive to a handful of engine and weather variables.